v1.5.3

Fixed

• The within-run abstraction dedupe shipped in 1.5.2 was set where it caught nothing. Its default of 0.82 was a guess. Measured on the live store that motivated it (qwen3-embedding:0.6b): the paraphrase pairs from one run score 0.615–0.693 against each other; a later run's five genuinely distinct abstractions top out at 0.539. Default is now 0.60, sitting between the two measured sets, and abstraction_dedupe_threshold is a namespace config key passed through by the dream tool, because other embedding models spread scores differently and the right line is a per-deployment measurement. (#83)